
Yeah Phil's a really nice guy (and for a time Eric Lafortune was there too), really helpful in getting the math part of it right. His global illumination compendium is invaluable.
You're right, I'm not doing bidirectional, I'm shooting from the camera with shadow rays & importance sampling. Sampling was done using QuasiMontecarlo (HaltonHammersly inverse radical sequence). Stratified sampling works well for lower dimensions but in higher dimensions takes a looooooong time to fill up. I think Pixar has patented stochastic sampling so if you want to use it in anything other than research it has to be QMC.
Also, I'm not simulating everything there is... things like diffraction, florescence, subsurface scattering, participating media etc aren't modeled. Besides, the input BRDF from max is hardly physically realistic. Most of the surfaces in the flamingo picture are either Lambertian or Lambertian with a delta function (mirror reflection). Other images do have gloss, but even then the BRDF is just plain olde modified Phong.
As a side note I would recommend using the Mersenne Twister for doing Montecarlo integration (again Pixar patents on stochastic sampling apply). The standard crand() breaks down in about 24 dimensions.
